Meta-learning for multimodal data

How can we transfer knowledge across tasks and domains to improve machine learning on multimodal data?

Learn more
Menu

Research areas

Time series High dimensional inference Evolution & adaptation Multi-agent systems Reinforcement learning Unsupervised learning Supervised learning Semi-supervised learning Deep learning Applications (Machine learning)

Introduction

This group aims to bring together machine learning researchers, data scientists, and domain experts from diverse backgrounds, career stages, and disciplines to develop algorithms and tools that transfer knowledge across tasks and domains to improve the performance of learning algorithms on data of multiple modalities in real-world applications. We call this problem meta-learning for multimodal data, adopting a broad definition of meta-learning to bring researchers and practitioners in related areas together.

Explaining the science

For example, in healthcare, there are data of multiple modalities including medical images, health-monitoring data, electronic health records, and multi-omics data. In practice, clinicians often make decisions using data from more than one modality and leveraging experience from related tasks/domains. To advance AI for such real-world problems, we aim to collaboratively develop meta-learning algorithms and tools that can leverage experience and knowledge on individual modalities, domains, and tasks to tackle real-world challenges in analysing data of multiple modalities across multiple domains for various tasks.

Aims

Meta-learning, also known as “learning to learn”, studies how to transfer knowledge on previous tasks and domains to improve the performance of learning algorithms on new tasks and domains, tackling the generalisation challenges slowing down wider adoption of AI and machine learning technologies. By adopting such a broad definition and focusing on data of multiple modalities in real-world applications, this interest group will bring people together sharing methodologies across multiple disciplines to generate innovative research ideas and facilitate nascent field technologies with applications across many disciplines of the Turing.

The main objectives of this interest group are:

  • To bring together researchers to develop innovative and practical algorithms as well as accessible and sustainable open-source software tools to advance research on meta-learning for multimodal data and tackle real-world challenges, e.g. in healthcare.
  • To establish an engaging community of researchers from multiple disciplines, including multimodal learning, transfer learning, domain adaptation, and data integration, and create more opportunities for early career researchers (ECRs) to take leadership roles and power future growth.
  • To create accessible materials suitable for dissemination to non-researchers and the general public, including online courses, tutorials, podcasts, and blogs.

Talking points

How can we develop innovative and practical models, algorithms, and software tools that can transfer knowledge across tasks and domains to improve the performance of machine learning on data of multiple modalities in real-world applications?

Motivation: This chosen topic aims to address the generalisation challenge in machine learning, a significant barrier in applications of AI in real-world problems. We define meta-learning broadly to enable the generation of new approaches and collaborative activities.

Expected outcomes: research publications, including review or position papers, and software tools that advance meta-learning for multimodal data.

How can we grow an engaging community of researchers on the topic above and create more opportunities for early career researchers?

Motivation: There is a critical shortage of researchers in machine learning and AI, particularly in academia. We need to take a community-based approach to ensure a sustainable talent pool, via encouraging the formation of local/regional communities and empowering more early career researchers to step up and take a leadership role.

Expected outcomes: A group of collaborative researchers, particularly early career researchers, taking a leading role in the development of theories and methodologies around meta learning for multimodal data.

How can we disseminate our research to non-researchers and particularly the general public so that they can better understand and support our research?

Motivation: AI ethics will have growing effects on the success/failure of AI technologies. To develop responsible, trustworthy, and deployable AI technologies, we need to disseminate our research and stimulate discussions beyond academia and involve the public in early stages of our research to ensure the relevance and acceptability of our research outputs.

Expected outcomes: Increased public awareness, understanding, and engagement on our research through online courses, tutorials, podcasts, blogs, as well as events open to public participation.

How to get involved

Click here to request sign-up and join

Recent updates

Professor Haiping Lu wrote a blog in Medium summarising the outcomes of the First Workshop on Multimodal AI organised by the Interest Group on 27 June 2023.

Read more.

Organisers

Shuo Zhou

Academic Fellow in Machine Learning, University of Sheffield

Contact info

[email protected]

We have also created a GitHub organisation for this interest group, with a homepage to enable more engaging and collaborative participation from the wider community.